add yolov3_mobilenet_v1

examples/yolov3_mobilenet_v1/configs/coco.yaml

@@ -0,0 +1,24 @@
+# COCO 2017 dataset http://cocodataset.org
+
+data:
+  dataset_name: coco
+
+  train_set: ./coco/train2017.txt  # 118287 images
+  val_set: ./coco/val2017.txt  # 5000 images
+  test_set: ./coco/test-dev2017.txt  # 20288 of 40670 images, submit to https://competitions.codalab.org/competitions/20794
+
+  nc: 80
+
+  # class names
+  names: [ 'person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', 'train', 'truck', 'boat', 'traffic light',
+           'fire hydrant', 'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog', 'horse', 'sheep', 'cow',
+           'elephant', 'bear', 'zebra', 'giraffe', 'backpack', 'umbrella', 'handbag', 'tie', 'suitcase', 'frisbee',
+           'skis', 'snowboard', 'sports ball', 'kite', 'baseball bat', 'baseball glove', 'skateboard', 'surfboard',
+           'tennis racket', 'bottle', 'wine glass', 'cup', 'fork', 'knife', 'spoon', 'bowl', 'banana', 'apple',
+           'sandwich', 'orange', 'broccoli', 'carrot', 'hot dog', 'pizza', 'donut', 'cake', 'chair', 'couch',
+           'potted plant', 'bed', 'dining table', 'toilet', 'tv', 'laptop', 'mouse', 'remote', 'keyboard', 'cell phone',
+           'microwave', 'oven', 'toaster', 'sink', 'refrigerator', 'book', 'clock', 'vase', 'scissors', 'teddy bear',
+           'hair drier', 'toothbrush' ]
+
+  train_transforms: []
+  test_transforms: []

examples/yolov3_mobilenet_v1/configs/hyp.scratch.yaml

@@ -0,0 +1,42 @@
+__BASE__: [
+  './coco.yaml',
+]
+
+epochs: 270
+per_batch_size: 8 # 8 * 8 = 64
+anchors: [[[116,90], [156,198], [373,326]], [[30,61], [62,45], [59,119]], [[10,13], [16,30], [33,23]]]
+
+optimizer:
+  optimizer: momentum
+  lr_init: 0.001  # initial learning rate (SGD=1E-2, Adam=1E-3)
+  gp_weight_decay: 0.0005  # optimizer weight decay 5e-4
+  momentum: 0.9  # SGD momentum/Adam beta1
+  nesterov: True # update gradients with NAG(Nesterov Accelerated Gradient) algorithm
+  loss_scale: 1.0 # loss scale for optimizer
+  warmup_epochs: 0  # warmup epochs (fractions ok)
+  warmup_momentum: 0.9  # warmup initial momentum
+  warmup_bias_lr: 0.0  # warmup initial bias lr
+  min_warmup_step: 4000 # minimum warmup step
+  group_param: yolov3 # group param strategy
+  start_factor: 1.0
+  end_factor: 0.01
+
+loss:
+  ignore_thresh: 0.7
+  downsample: [32, 16, 8]
+  label_smooth: false
+
+data:
+  num_parallel_workers: 4
+
+  train_transforms:
+    - { func_name: mixup, alpha: 1.5, beta: 1.5 }
+    - { func_name: random_distort }
+    - { func_name: random_expand, fill_value: [123.675, 116.28, 103.53]}
+    - { func_name: random_crop }
+    - { func_name: fliplr, prob: 0.5 }
+
+  test_transforms:
+    - { func_name: resize, target_size: [608, 608] }
+    - { func_name: normalize_image, mean: [0.406, 0.456, 0.485], std: [0.225, 0.224, 0.229] }
+    - { func_name: image_transpose, bgr2rgb: True, hwc2chw: True }

examples/yolov3_mobilenet_v1/group_params.py

@@ -0,0 +1,81 @@
+import numpy as np
+
+
+__all__ = ["create_group_param"]
+
+
+def create_group_param(params, gp_weight_decay=0.0, **kwargs):
+    """
+    Create group parameters for optimizer.
+
+    Args:
+        params: Network parameters
+        gp_weight_decay: Weight decay. Default: 0.0
+        **kwargs: Others
+    """
+    if "group_param" in kwargs:
+        return group_param_yolov3(params, weight_decay=gp_weight_decay, **kwargs)
+    else:
+        return params
+
+
+
+def group_param_yolov3(
+    params,
+    weight_decay,
+    start_factor,
+    end_factor,
+    lr_init,
+    warmup_bias_lr,
+    warmup_epochs,
+    min_warmup_step,
+    accumulate,
+    epochs,
+    steps_per_epoch,
+    total_batch_size,
+    **kwargs
+):
+    # old: # weight, gamma, bias/beta
+    # new: # bias/beta, weight, others
+    pg0, pg1, pg2 = _group_param_common3(params)
+
+    lr_pg0, lr_pg1, lr_pg2 = [], [], []
+    lrs = []
+    lrs.extend([lr_init] * 216 * steps_per_epoch)
+    lrs.extend([lr_init * 0.1] * 27 * steps_per_epoch)
+    lrs.extend([lr_init * 0.01] * 27 * steps_per_epoch)
+
+    warmup_steps = max(round(warmup_epochs * steps_per_epoch), min_warmup_step)
+    xi = [0, warmup_steps]
+    for i in range(epochs * steps_per_epoch):
+        _lr = lrs[i]
+        if i < warmup_steps:
+            lr_pg0.append(np.interp(i, xi, [warmup_bias_lr, _lr]))
+            lr_pg1.append(np.interp(i, xi, [0.0, _lr]))
+            lr_pg2.append(np.interp(i, xi, [0.0, _lr]))
+        else:
+            lr_pg0.append(_lr)
+            lr_pg1.append(_lr)
+            lr_pg2.append(_lr)
+
+    nbs = 64
+    weight_decay *= total_batch_size * accumulate / nbs  # scale weight_decay
+    group_params = [
+        {"params": pg0, "lr": lr_pg0},
+        {"params": pg1, "lr": lr_pg1, "weight_decay": weight_decay},
+        {"params": pg2, "lr": lr_pg2},
+    ]
+    return group_params
+
+
+def _group_param_common3(params):
+    pg0, pg1, pg2 = [], [], []  # optimizer parameter groups
+    for p in params:
+        if "bias" in p.name or "beta" in p.name:
+            pg0.append(p)
+        elif "weight" in p.name:
+            pg1.append(p)
+        else:
+            pg2.append(p)
+
+    return pg0, pg1, pg2  # bias/beta, weight, others